Providing Hand Use Context for Outpatient Neurorehabilitation with Egocentric Object Detection
Adesh Kadambi; Jose Zariffa
Abstract
Recent advancements in wearable technology and machine learning have the potential to enhance rehabilitation therapy, particularly in outpatient settings. However, to effectively support therapy planning, such technologies need to capture context-specific information about an individual's activities of daily living (ADLs). In this study, we evaluated the performance of two object detection models, Detic and UniDet, on egocentric videos recorded by individuals with spinal cord injury (SCI). Our evaluations revealed that UniDet, when evaluated on its original 700 classes, achieved a Mean Average Precision (mAP) of 0.0382 for all objects and 0.0988 for active objects. When evaluated on a set of 27 consolidated functional categories, UniDet's performance improved to an mAP of 0.1503 for all objects and 0.1910 for active objects. Detic demonstrated superior performance with an mAP of 0.1772 for all objects and 0.2754 for active objects when evaluated on the 27 functional categories. However, the ground truth labelling strategy resulted in a large number of false positives, suggesting that the model performance is likely higher. Despite challenges posed by low-light conditions and motion blur, this study provides crucial insights into the potential of object detection models in therapy planning, facilitating the integration of wearable technology and machine learning in outpatient rehabilitation and enabling more personalized and effective therapeutic strategies.
Keywords: egocentric video; object detection; spinal cord injury; neurorehabilitation; wearable technology
Links
[Full text PDF][Bibtex]